The Last Mile Is Where Contact-Center AI Gets Real

Host Sonya Huang opens with the labor thesis that hangs over every AI pitch: contact centers are a giant pool of work begging to be automated, and the TAM is measured in the tens of trillions. She asks Cresta CEO Ping Wu a blunt question — if he had to guess, how much of call-center labor spend will actually be fully automated by AI? Wu, who built Google's contact center business before becoming product leader and then CEO at Cresta, does not pretend the firm has a private answer.

Internally, Cresta has spent a lot of time debating it. Wu says the honest answer is that nobody knows for sure, and the number you hear depends on what the speaker is selling. Some people claim 100 percent of humans will be gone. Gartner research, he notes, shows that none of the Fortune 500 will have a fully humanless contact center over the next five years. When GPT-4 first came out, a lot of people said humans would be gone in two or three years. Cresta's belief at the time was that transformation, especially inside existing Fortune 500 companies, would take far longer than that.

Doug Leone, the Sequoia partner who sits on Cresta's board, draws a sharper distinction. At the limit, it is 100 percent. He is also mindful that IBM mainframes and COBOL still run in the American banking system. So the percentage is not the metric he cares about. Whether the answer is 30 or 60 percent, a fifty-year shift means one thing for a company like Cresta, and a three-year shift means something else. The n-number is not relevant. Speed of adoption is.

Wu starts with a definition. People hear "call center" and picture humans answering phones. Contact center is broader: omnichannel work across email, digital chat, websites, apps, and voice. Historically about 17 to 20 million human agents work in it. The software market is in the tens of billions of dollars; some research puts the AI market in the high tens of billions. And it is not only complaints. Roughly a quarter of the work is revenue-generating — selling, collecting money, retaining customers.

Leone dates himself. Thirty years back, names like Avaya drifted in and out of bankruptcy. Fifteen years back it was Genesys and peers. Why would a bright young engineer, Ping Wu of a decade-plus ago, walk into a stodgy, slow-growing market of low interest? Wu admits luck. Fifteen years ago he did not even realize that long history of slow growth; if he had, he might have thought differently. At the time, conversational AI excitement clustered around consumer speakers that people thought would disrupt Google and become the entry point for all consumer interaction. Wu bet the other way: the contact center would be the most exciting place for conversational AI to transform.

The reasons were the ones venture capitalists usually get excited about, plus one more. It is a massive market, full of humans, sitting in the middle between businesses and customers, with all interaction flowing through it. And no one is happy. Customers wait too long. Agent attrition averages 35 to 40 percent; during Covid some companies had more than 100 percent turnover, meaning they replaced the entire workforce in a year, in a job where people get yelled at all day. Businesses always feel they could do more with less. Wu's one-word prescription is abundance: technology that makes all three parties better off.

Technology has already hit this industry in waves. First came IVR: press one for billing, press two for a outage. Then input got better — speak natural language instead of pressing buttons — as NLP and text-to-speech improved. When Wu started contact-center AI at Google, it was even before BERT and before transformers. Models did intent classification and entity extraction. The conversation itself was still manually crafted. Transformers arrived and were first used the same way, still for classification, still with hand-built flows. LLMs changed both sides: they can run the conversation, and they can understand it with a depth that was not possible before.

Wu would rather start from first principles than from a happiness score for each generation of tech. A lot of conversations should not happen at all. The fact that they do usually means a process is broken, a website update freaked people out, or a firmware update took down a network. Use AI to bring 100 percent visibility into every interaction, analyze it, find the root cause, and fix that first. Beyond that, automate the low-emotion-value interactions that neither the business nor the customer wants — those should be self-serve. On top, contact-center AI should enable conversations companies cannot afford to staff today.

Huang asks whether end customers will ever prefer an AI agent to a human, and whether that point has arrived. Wu has never met anyone who, frustrated on a support call, said "send me your AI, please." Search a company's customer-service pages and the first question Google surfaces is almost always how to talk to a live person. Huang pushes back from the robot side: AI has infinite patience, is not trying to hit time-to-resolution, and can be more empathetic than a human having a bad day. Leone reaches for gold versus Bitcoin. There is a language component, a training component, and a human-patience component, and he thinks AI wins on all of them in the next two to three years. Bitcoin will be worth more than gold, he adds — not investment advice.

Cresta's playbook is to meet customers where they are. Self-driving only has economic impact if you automate 100 percent of the time. Contact-center work is different: every conversation is an independent unit. You can automate the share that is ready, and on the rest you can still use AI to assist humans — take the first 10 percent (authentication, intake, lead qualification), take all the after-call work, and sit in the middle of the live conversation to retrieve knowledge and enter data. Those are not mutually exclusive. Until a customer is ready to turn the contact center fully over to AI, the journey takes different amounts of time depending on the business and the IT stack.

Where customers sit today depends on complexity. If Huang and Wu opened an e-bike store on Shopify this afternoon, they could automate 100 percent, because the product is simple. A company that touches many countries and tens of millions of people is an order of magnitude harder. The other constraint is infrastructure. A lot of what humans do in contact centers is optimized for humans: systems of record and ticketing that have been around for decades, many with no APIs. The only way to change a record is through a GUI. Those are not AI problems. Cresta works with customers to build the real-time APIs, which is why the timeline varies so much by business.

Huang reaches for Tesla. Part of the path to full autonomy is the data coming off cars even at Level 2. Because Cresta sits in agent-assist, it gets the full conversation — voice and digital — which can become the training base for automating more over time. Wu agrees, and says the lesson was counterintuitive. He joined seven or eight years ago believing in automation only. Real deployments broadened his horizon: to do the best possible automation, you need to know what humans are actually doing, including what they see on the screen, not just the transcript. Leone puts the product split in plainer language. One side is the sizzle everyone wants to talk about on day one. The other is the operational reality of running a contact center. New-age companies without data and systems hit a wall. Companies with only the operational product get labeled tired. Cresta doubled down on both.

Customer support is a consensus category. Everyone has the same LLMs and the same vision. Wu's competitive answer starts with what a model is: a bunch of weights and data that, by itself, does not provide value. The question is how much you have to build on top to deliver that value. If the layer is very thin, you do not have much chance to accrue value. If the layer disappears when the model gets better, you do not have a durable business. Contact centers are not that case. A majority of agencies are still on-premise. Fortune 500 agents, in surveys Cresta has seen, interact with eight to ten different systems. Those companies acquired other companies over decades; backends may not talk to each other. Book a flight in one system, a hotel in another.

Huang asks what is overhyped and underhyped. Overhyped, Wu says, is the scarcity mindset — short-term job displacement. Underhyped is abundance: new experiences AI can enable. Can you talk to a website? Directly to an app? Turn a synchronous interaction asynchronous — tell an airline app to do X, Y, and Z and call you back when it is done? Multilingual agents can run conversations that do not happen today simply because there is no staff. He also thinks people are obsessed with the workforce side of the ledger. Everyone asks how many agents AI will replace. Almost nobody asks how many inbound calls AI will replace. Over the next few years he expects a race to put AI assistants on consumer aggregators, and consumers will dedicate a lot of tasks to those assistants, including making the phone calls.

Huang later asks what separates a fleshy demo from production. Wu says AI has a unique gap that rocket launches do not: for a rocket, the demo is the production, and you cannot fake it. Auto-summary looks like a commodity anyone can ship with ChatGPT. Deploy it in a 20,000-person contact center across continents and the list explodes. Getting real-time audio is easy in a cloud demo; about 50 percent of conversations still happen on-premise, and access can cost a lot of money. Real calls have transfers and third-party specialists who also need to be transcribed. Some calls run three or four hours, past the context window. There is background noise, different templates by call reason that you cannot miss, PII that cannot sit at rest, and data residency for multinational banks and healthcare providers. Leone's punchline: that is why you need a product-minded CEO. Huang's: all the pain and all the value is in the last mile, which is why value sits in the application layer.

Huang asks Leone how building a company in AI differs from the last few decades of building legendary ones. His answer is that it is not very different. You need a terrific founder. You need to plug in world-class engineers at the start: unless you start with A-pluses, you never move up, you only move down. You hire salespeople who are not administrators — maybe they were regional managers early on — because you cannot get world-class sellers, and if you did they would be too big for the company. You have to know the ramp you are willing to fund, the role of marketing, and what Leone calls the merchandising cycle: product marketing to BDRs to revenue. Wherever that is broken, it looks like a bad VP of sales.

Speed is the part he almost folded into that answer. Every board he is on moves with extreme speed, because he paints founders a river with rocks. The founder's and CEO's job is to remove those rocks. When they bring next year's plan, he does not care that it is 150 percent net-new ARR growth. He wants to know why the plan is the plan, and he wants to challenge them on why it is not 3x that. Funding is rarely the constraint in this market. Management experience often is. Market is not an excuse a little company gets to use. He also wants the ramp linear, so the company can make mid-course corrections. Hire 250 salespeople in Q1, discover a product problem in Q3, and you are stuck with the burn. "We have 10 fingers, 100 percent growth" is baloney. The mantra on every board he has served: how fast can we possibly grow? AI is not different.

What does Cresta need over the next five-plus years? Keep putting one foot in front of the other on product. When people hit the Peter principle, be aggressive about hiring the people who can take it from there, and stay away from "very experienced" people who start feeling like suits and administrators. The other gap is marketing. A lot of companies have sizzle and no steak. Cresta has a bunch of steak: best-in-class in agent assist, heading that way in the automated agent, with a growing run rate in both. It needs a marketing overlay so it becomes a household name. Huang's earlier summary still fits: vertical integration from the steak to the sizzle is how you win.

Are we in an AI bubble? Leone unpacks the word. Bubble implies you invest money and lose it, from too few companies or too much capital. There is certainly an abundance of capital. What he has noticed across cycles is shrinking latency from technology birth to a real market. Netscape went public in 1995; Google and Amazon were built in the late 1990s; then came a pause when people said the internet was a fraud, and three years later the world went crazy. In mobile the delay was shorter. Former Sequoia partner Jim Goetz looked at $19 apps and asked how you build a multi-billion-dollar company, never thinking of Airbnb or DoorDash — which showed up a year or two later. AI has shrunk that latency further. You have to invest. You are at the front end of a cycle, which does not mean you invest in everything. One Sequoia mistake: when a small company shows revenue momentum at the front of a market, someone in the partners meeting says it can stop or be substituted. Leone's rule is to lean in and hold your nose on price.

Where does value occur — compute, infrastructure, foundation models, or applications? "Up." It always accrues up; look at gross margins as you move up the stack from chips to systems. Sequoia was Nvidia's first investor, and Jensen Huang seeing the future years ahead was, in Leone's telling, probably the greatest coup in Silicon Valley. Over time, though, he thinks value accrues to the application layer, near the customer, near the money, near the business user. Internet and mobile were tools that made us networked and mobile. He views the AI wave as Industrial Revolution 2.0, much larger. Five years ago people thought they had already seen the biggest market caps, because connectivity created that revenue growth. Nobody imagined something bigger than connectivity: a complete redoing of how humanity exists, works, lives, and enjoys. AI will be a wonderful thing, and maybe a kiss of death, over the next ten to twenty years.

Wu agrees, and adds the thing that makes AI unlike those earlier waves: surprise. Take a 2015 worldview and hand it to someone in 2007 when Steve Jobs introduced the iPhone, and they could resonate with it. The same for the internet. As models get better, capabilities appear that even the authors of the transformer paper would not have imagined. A lot of the improvement is nonlinear — zero-to-one, still happening at the bottom layer. Leone closes with a story from March 2022, his last Sequoia annual meeting. He showed a slide of waves: chips, systems, LAN/WAN, internet, mobile, and then a question mark. Sequoia is as close to the bleeding edge as anyone in seed, and the partnership did not see the wave coming. "This wave has been a tsunami and I don't think there's any end in sight."

Leone has to leave early — he is in the middle of recoding some things, and when Huang asks if he is vibe coding, he says yes, he is vibe coding everything. Wu walks through the stack. The voice AI agent streams bidirectional audio end to end and orchestrates a chain: speech-to-text, noise cancellation, turn-taking and interruption handling, a foundation model for the dialogue, text-to-speech on the way out, plus smaller models in parallel for guardrails (never give tax advice, never make financial promises). Across the platform there are about 20 models, some open source and fine-tuned, some tiny type-ahead models for chat and email. TTS uses ElevenLabs and other vendors, compared constantly. Latency is below 800 milliseconds, which is why it feels like talking to a human. The agent-assist path is similar but one-directional: listen to two humans, orchestrate ten-plus models, and let customers train custom event detectors — Cresta calls the tool Opera — distilled via teacher-student into tiny real-time models for fraud, objections, and workflow triggers.

Controlling the conversation without sliding back into IVR is, Wu says, how you train a new employee: give a spec for the goal and the tools, and let the LLM handle messy, nonlinear human talk. Anything you can write down in code, step by step, is a workflow — a car wash, a boba-tea line. Human conversation is not that. Agentic behavior is the LLM's job; determinism comes from testing, large-scale simulation, and guardrails so a change anywhere in the system still produces expected behavior. Cresta maps human calls into conversation blueprints, using the model to surface unknown unknowns and the 577 different ways people express the same intent. Those traces feed a visitor model for simulation. The firm uses prompting, RAG, and fine-tuning, and is exploring reinforcement learning that ties human behavior to outcomes. For digital chat, the foundation model is already pretty good if you get the most out of it; summarization and autocomplete still have a lot of room in fine-tuned open-source models.

If everything goes right, Wu thinks AI will disappear the way electricity did, into workflows. Twenty or thirty years later, no one will realize whether they are talking to AI or to a human assisted by AI. What excites him is curing the company's multiple personality. In the sales and marketing phase, businesses court you aggressively. Once you sign, you meet an entirely different personality: service departments talking about tier defense and deflection, treating the person they were courting days ago as someone to fend off. You spend half an hour on the support line sharing feedback, and two weeks later another department sends a survey asking for the same thing. AI agents can make the entire customer journey one continuous conversation, and the LLM is a perfect tool for that — personalization that was not possible before. Huang's lasting note is the scarcity-versus-abundance frame: how far business-to-customer communication can evolve if you bring abundance into the field.